The finding
Google Research's systematic evaluation of 180 AI agent configurations reveals that multi-agent systems excel at parallelizable tasks but actively harm performance on sequential workflows. Centralized coordination delivered an 80.8% boost on parallel tasks, while independent agents degraded sequential task performance by 39-70%.
The team tested five architectures—single-agent, independent, centralized, decentralized, and hybrid—across three LLM families (GPT, Gemini, Claude) using four benchmarks: Finance-Agent (financial reasoning), BrowseComp-Plus (web navigation), PlanCraft (game planning), and Workbench (workflows).
What matters for CTOs
Two practical implications:
First, the predictive model (R²=0.513-0.524) correctly identifies optimal architecture for 87% of held-out tasks using task properties like decomposability and tool density. This means organizations can match agent designs to workload characteristics rather than defaulting to multi-agent systems.
Second, error amplification is real. Independent agents showed 17.2x error multiplication; centralized systems 4.4x. Under fixed computational budgets, multi-agent coordination overhead outweighs benefits once single-agent baseline performance exceeds roughly 45%—a capability saturation point most vendors don't discuss.
The broader context
This challenges the "more agents always better" heuristic that's driven enterprise AI architecture decisions since papers like "More Agents Is All You Need" emerged. The research suggests agent scaling follows task-specific principles, not universal laws—similar to how distributed systems don't automatically improve every workload.
The predictive model validated on GPT-5.2 with MAE=0.071, confirming four of five principles generalize across model generations. Worth noting: all data comes from Google authors; no independent replication yet.
Trade-offs
Decentralized peer-to-peer coordination added only 9.2% on web navigation—modest returns for significantly higher communication overhead. The study used fixed computational budgets, meaning every agent added reduces per-agent resources. In production, this translates to cost-performance calculations most enterprises haven't modeled.
The research provides quantitative guardrails for agent system design, but implementation depends on accurate task characterization—itself a non-trivial problem at enterprise scale.